Elamur Research⌕
☰
home / papers / 2009 / Improving the Arabic Pronunciation Dictionary for Phone and Word Recognition with Linguistically-Based Pronunciation Rules

Improving the Arabic Pronunciation Dictionary for Phone and Word Recognition with Linguistically-Based Pronunciation Rules

Paper · 2009 · Fadi Biadsy, Nizar Habash and Julia Hirschberg

Summary

The paper addresses a weak point in Modern Standard Arabic (MSA) speech recognition: the pronunciation dictionary. Arabic script usually omits diacritics, and the actual sound of a word depends on morphology and context. Examples are sun-letter assimilation of the definite article, the silent or pronounced Ta-Marbuta, Hamzat-Wasl and silent Alifs. Earlier systems turned each diacritized word into one pronunciation using only a handful of simple rules. The authors instead run MADA (morphological analysis and disambiguation) on the text and then apply a set of phonology-based rules. These rules map each diacritized form to one or more plausible pronunciations, including variants that speakers commonly produce even when they are not "correct." From this output they build two separate lexicons, one for training acoustic models and one for decoding. Compared with a re-implemented baseline from Vergyri et al. (2008), their systems (XPR, XWR) raise phone recognition accuracy by 3.77–7.29% absolute and word recognition accuracy by 4.1% absolute. This shows that explicit linguistic knowledge in the lexicon pays off directly in ASR.

Key points

  • The rules fall into four groups: (I) shared with the baseline, (II) modified baseline rules such as Shadda and Alif-Maqsura, (III) new rules such as Sun letters, the definite article and Hamzat-Wasl, and (IV) rules that generate extra variants. For each rule, the paper reports what fraction of the 335,324-word training corpus it affects; the definite article rule touches 30.0%.
  • Word recognition accuracy rises from 52.78 (BASEWR) to 56.88 (best XWR) on 36,538 test words. Using rule groups I–III alone gives +2.27, and adding group IV gives a further +1.24.
  • The best setup uses only MADA's top analysis in the training dictionary but the top two analyses in the decoding dictionary. The second analysis is added only when the confidence gap between the two is below a threshold.
  • The lexicons grow from 1 pronunciation per word (baseline training dictionary) to 3.05 (XWR decoding dictionary with top-1 and top-2 analyses, 105,402 entries).
  • Setup details:
    • Language model: trigram LM built with SRILM on about 281M words (training transcripts plus a subset of Arabic Gigaword).
    • Vocabulary: 34,511 undiacritized words, with a test OOV rate of 6.38%.
    • The dictionary-generation script was released publicly.

Interesting ideas

  • I1. The system keeps two lexicons with different breadth: a narrow single-best one for training, so the models learn sharp distinctions, and a broader multi-variant one at decoding time, so that real-world variation is tolerated.
  • I2. A second analysis is admitted only when the disambiguator's confidence margin between its top two choices is small, so ambiguity is added only where it is real.
  • I3. Each hand-written rule is reported together with how much of the training data it changes, which makes the rule set auditable and shows which rules carry the most weight.
  • I4. Variant-generating rules are ablated as a separate group (I–III versus all), which isolates the value of producing extra alternatives from the value of correcting single outputs.
  • I5. As a robustness check, the authors drop noisy utterances where the baseline does badly, which deliberately favours the baseline, and show their system still wins.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Train the parser and tagger on single-best silver analyses, but let inference consider a small k-best set of lemma/tag readings that later stages can resolve.
  • I2 (tokens): The tagger and lemmatizer could emit an alternative reading only when the score gap is narrow, keeping downstream frames compact while staying explainable about genuine ambiguity.
  • I3 (learn): Self-induced repair rules could be logged with their coverage on the treebank, so low-impact or suspicious rules are easy to spot and prune.
  • I4 (learn): Evaluate "add alternatives" components separately from "fix the single answer" components, so each gains or loses credit on its own.
  • I5 (answer): When reporting reader gains, also test on a subset filtered in the baseline's favour, as a negative-control style check that the improvement is not an artifact of the test data.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Columbia UniversityNew York · United States241 works · 54,053 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksSpeech & spoken language · 480 works

Cite

Cite as Fadi Biadsy, Nizar Habash and Julia Hirschberg. 2009. Improving the Arabic Pronunciation Dictionary for Phone and Word Recognition with Linguistically-Based Pronunciation Rules. Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{biadsy2009improving,
  title = {{Improving the Arabic Pronunciation Dictionary for Phone and Word Recognition with Linguistically-Based Pronunciation Rules}},
  author = {Fadi Biadsy, Nizar Habash and Julia Hirschberg},
  booktitle = {Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
  doi = {10.3115/1620754.1620812},
  url = {https://doi.org/10.3115/1620754.1620812}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M. Afify, L. Nguyen, B. Xiang, S. Abdou, and J. Makhoul. 2005. Arabic broadcast news transcription using a one million word vocalized vocabulary. In Proceedings of Interspeech 2005, pages 1637–1640. find
  2. S. Ananthakrishnan, S. Narayanan, and S. Bangalore. 2005. Automatic diacritization of arabic transcripts for asr. In Proceedings of ICON, Kanpur, India. find
  3. F. Biadsy, J. Hirschberg, and N. Habash. 2009. Spoken Arabic Dialect Identification Using Phonotactic Modeling. In Proceedings of EACL 2009 Workshop on Computational Approaches to Semitic Languages, Athens, Greece.
  4. T. Buckwalter. 2004. Buckwalter Arabic Morphological Analyzer Version 2.0. Linguistic Data Consortium, University of Pennsylvania, 2002. LDC Cat alog No.: LDC2004L02, ISBN 1-58563-324-0. find
  5. Y. A. El-Imam. 2004. Phonetization of Arabic: rules and algorithms. In Computer Speech and Language 18, pages 339–373. ↗
  6. N. Habash and O. Rambow. 2005. Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 573–580.
  7. N. Habash and O. Rambow. 2006. MAGEAD: A Morphological Analyzer and Generator for the Arabic Dialects. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 681–688, Sydney, Australia. ↗
  8. N. Habash and O. Rambow. 2007. Arabic Diacritization through Full Morphological Tagging. In Proceedings of the 8th Meeting of the North American Chapter of the Association for Computational Linguistics/Human Language Technologies Conference (HLT-NAACL07).
  9. N. Habash, A. Soudi, and T. Buckwalter. 2007. On Arabic Transliteration. In A. van den Bosch and A. Soudi, editors, Arabic Computational Morphology: Knowledge-based and Empirical Methods. Springer. ↗
  10. M. Maamouri, A. Bies, H. Jin, and T. Buckwalter. 2003. Arabic treebank: Part 1 v 2.0. Distributed by the Linguistic Data Consortium. LDC Catalog No.: LDC2003T06. ↗
  11. A. Messaoudi, J. L. Gauvain, and L. Lamel. 2006. Arabic broadcast news transcription using a one million word vocalized vocabulary. In Proceedings of ICASP 2006, volume 1, pages 1093–1096. find
  12. H. Soltau, G. Saon, D. Povey, L. Mangu, B. Kingsbury, J. Kuo, M. Omar, and G. Zweig. 2007. find
  13. The IBM 2006 GALE Arabic ASR System. In Proceedings of ICASP 2007. find
  14. S. Stolcke. 2002. Tokenization, Morphological Analysis, and Part-of-Speech Tagging for Arabic in One Fell Swoop. In Proceedings of ICSLP 2002, volume 2, pages 901–904. find
  15. D. Vergyri and K. Kirchhoff. 2004. Automatic Diacritization of Arabic for Acoustic Modeling in Speech Recognition. In Ali Farghaly and Karine Megerdoomian, editors, COLING 2004 Workshop on Computational Approaches to Arabic Script-based Languages, pages 66–73, Geneva, Switzerland. ↗
  16. D. Vergyri, A. Mandal, W. Wang, A. Stolcke, J. Zheng, M. Graciarena, D. Rybach, C. Gollan, R. Schluter, K. Kirchhoff, A. Faria, and N. Morgan. 2008. Development of the SRI/Nightingale Arabic ASR system. In Proceedings of Interspeech 2008, pages 1437–1440. find
  17. S. Young, G. Evermann, M. Gales, D. Kershaw, G. Moore, J. Odell, D. Ollason, D. Povey, V. Valtchev, and P. Woodland. 2006. The HTK Book, version 3.4: htk.eng.cam.ac.uk. Cambridge University Engineering Department. find
  18. I. Zitouni, J. S. Sorensen, and R. Sarikaya. 2006. Maximum Entropy Based Restoration of Arabic Diacritics. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 577–584, Sydney, Australia.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/biadsy-2009-improving-arabic-pronunciation-dictionary-phone/